Open Weights vs. Open Source AI: What's the Difference?

"Open" gets used loosely across AI model releases, and it covers a wide range of what's actually available.

Open weights

This is the narrowest common meaning: the trained parameter files themselves are publicly downloadable and usable, often for both research and commercial purposes. It says nothing about whether the training code, training data, or full training pipeline is public -- most "open weight" releases keep those private.

Open source AI

The Open Source Initiative (OSI), the organization that maintains the canonical Open Source Definition for software, has worked on a formal definition of open source AI that goes further than weight availability -- addressing whether enough information about training data, code, and parameters is available for the system to be meaningfully studied, modified, and reproduced. By that fuller standard, many popular "open" model releases don't actually qualify, since their training data and pipelines remain undisclosed.

Licensing is a separate axis entirely

Even among open-weight releases, license terms vary widely -- from permissive licenses like Apache 2.0 or MIT, to custom community licenses that restrict commercial use above a certain scale or require specific attribution. "Open weights" doesn't imply "no restrictions."

Where hashing and timestamping fit in

Weight Registry is deliberately orthogonal to this whole debate. Whether a model is "truly" open source by any particular definition is a separate question from whether you can verify exactly which bytes were released under whatever license was stated, and when. Hashing and on-chain timestamping answer the second question regardless of how the first one is settled.